Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RTF: Region-based Table Filling Method for Relational Triple Extraction

Ning An, Lei Hei, Yong Jiang, Weiping Meng, Jingjing Hu, Boran Huang, Feiliang Ren

TL;DR

This work presents Region-based Table Filling (RTF) for relational triple extraction, addressing the spatial gaps of token- and token-pair–level methods. By introducing the Entity Pair as Region (EPR) tagging scheme, region-level representations via a Convolution Block, relational residual learning, and a bi-directional decoding process, RTF effectively identifies entity-pair regions on relation-specific tables. Empirical results on NYT and WebNLG (including rearranged variants) demonstrate state-of-the-art performance and enhanced generalization, with notable recall improvements driven by regional correlations. The approach advances boundary recognition and overlap handling, offering a scalable method for robust RTE in knowledge-graph construction, albeit with higher memory and training-time demands for large-scale, multi-relation settings.

Abstract

Relational triple extraction is crucial work for the automatic construction of knowledge graphs. Existing methods only construct shallow representations from a token or token pair-level. However, previous works ignore local spatial dependencies of relational triples, resulting in a weakness of entity pair boundary detection. To tackle this problem, we propose a novel Region-based Table Filling method (RTF). We devise a novel region-based tagging scheme and bi-directional decoding strategy, which regard each relational triple as a region on the relation-specific table, and identifies triples by determining two endpoints of each region. We also introduce convolution to construct region-level table representations from a spatial perspective which makes triples easier to be captured. In addition, we share partial tagging scores among different relations to improve learning efficiency of relation classifier. Experimental results show that our method achieves state-of-the-art with better generalization capability on three variants of two widely used benchmark datasets.

RTF: Region-based Table Filling Method for Relational Triple Extraction

TL;DR

This work presents Region-based Table Filling (RTF) for relational triple extraction, addressing the spatial gaps of token- and token-pair–level methods. By introducing the Entity Pair as Region (EPR) tagging scheme, region-level representations via a Convolution Block, relational residual learning, and a bi-directional decoding process, RTF effectively identifies entity-pair regions on relation-specific tables. Empirical results on NYT and WebNLG (including rearranged variants) demonstrate state-of-the-art performance and enhanced generalization, with notable recall improvements driven by regional correlations. The approach advances boundary recognition and overlap handling, offering a scalable method for robust RTE in knowledge-graph construction, albeit with higher memory and training-time demands for large-scale, multi-relation settings.

Abstract

Relational triple extraction is crucial work for the automatic construction of knowledge graphs. Existing methods only construct shallow representations from a token or token pair-level. However, previous works ignore local spatial dependencies of relational triples, resulting in a weakness of entity pair boundary detection. To tackle this problem, we propose a novel Region-based Table Filling method (RTF). We devise a novel region-based tagging scheme and bi-directional decoding strategy, which regard each relational triple as a region on the relation-specific table, and identifies triples by determining two endpoints of each region. We also introduce convolution to construct region-level table representations from a spatial perspective which makes triples easier to be captured. In addition, we share partial tagging scores among different relations to improve learning efficiency of relation classifier. Experimental results show that our method achieves state-of-the-art with better generalization capability on three variants of two widely used benchmark datasets.
Paper Structure (21 sections, 6 equations, 5 figures, 7 tables, 1 algorithm)

This paper contains 21 sections, 6 equations, 5 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Task Definition
  5. EPR Tagging Scheme
  6. RTF
  7. Table Representation
  8. Convolution Block
  9. Relational Residual Learning
  10. Bi-directional Decoding
  11. Loss Function
  12. Experiments
  13. Experiments Settings
  14. Main Results
  15. Analysis on Complex Scenarios
  16. ...and 6 more sections

Figures (5)

  • Figure 1: Example of our EPR tagging scheme and bi-directional decoding. This sentence contains Single Entity Overlapping (SEO) and Head Tail Overlapping (HTO) overlapping patterns.
  • Figure 2: RTF overview. Left: First, we construct a token pair-level table representation. Then capture the local dependencies and construct a region level table representation by Convolution Block. After that, we obtain the tagging score of each cell from the sharing score of all relations $\boldsymbol{z}_{ij}^e$ and the relation-specific residual $\boldsymbol{z}_{ij}^r$. Finally, we obtain triples by bi-directional decoding strategy. Right: Details of the Convolution Block. Similar to ResNet DBLP:conf/cvpr/HeZRS16, Conv Block consists of $1\times1, 3\times3, 1\times1$ convolution and residual connection.
  • Figure 3: Recall for two different tasks, entity pair recognition and triple extraction.
  • Figure 4: Interaction when combining convolution with EPR tagging scheme. From left to right is the scene when the convolution interacts with "UL", "BR" and "SP" respectively.
  • Figure 5: Visualization of Tagging Score. Brighter colors mean larger scores. Each column from left to right is the relation independent score $\boldsymbol{z}_{ij}^{e}$, the relation residual $\boldsymbol{z}_{ij}^{r}$ for "location" and "floorCount", respectively.